Catalog / MCP servers

DuckDB / MotherDuck MCP server

Run DuckDB SQL from an agent over local files, in-memory data, S3 or MotherDuck — good for quick analysis of CSV, Parquet and JSON without a warehouse.

Install

  1. Claude Code — local DuckDB file, read-only

    claude mcp add --scope user duckdb --transport stdio -- uvx mcp-server-motherduck@1.0.8 --db-path /absolute/path/to/db.duckdb
  2. Claude Code — in-memory scratch database

    claude mcp add --scope user duckdb --transport stdio -- uvx mcp-server-motherduck@1.0.8 --db-path :memory: --read-write --allow-switch-databases

    The vendor's "dev mode". No guardrails — it can read and write local files and switch to any database.

  3. Claude Desktop

    One-click .mcpb bundle from the repo's latest release.

MotherDuck’s local MCP server for DuckDB. Five tools: execute_query, list_databases, list_tables, list_columns, and (if enabled) switch_database_connection. DuckDB reads CSV, Parquet and JSON directly from disk or S3, which makes this a fast way to let an agent explore an export or a data dump.

When to use it

  • Ad-hoc analysis of files on your machine without loading them into a warehouse.
  • Prototyping queries before they go into dbt or a production pipeline.
  • Querying a MotherDuck cloud database from your editor.

Before you install

Since v1.0 the server is read-only by default for files and MotherDuck; --read-write turns writes on. In-memory databases are always writable. MotherDuck in read-only mode needs a read-scaling token. Don’t expose the HTTP transport to other people — MotherDuck recommends its hosted Remote MCP for that, or SaaS mode plus --init-sql hardening if you self-host.

Publisher
MotherDuck
Works with
Claude apps, Claude Code, Codex, Gemini, Cursor, Any agent
License
MIT
Cost
Free

Curated by Andrew Borden-Chisholm. Notes on building with these tools are on the blog.